Elyn Tan
AI & Technology|6 min read|September 2026

Malaysia Is Writing an AI Law. Accountability Has to Follow Control

Malaysia's proposed AI governance framework ties responsibility to each actor's role and degree of control. That principle will protect people only if the chain of control is visible before a system affects them, not reconstructed after harm occurs.

In brief

Malaysia has an opportunity to make AI accountability more than a principle. For material and high-risk systems, the law should make responsibility traceable across developers, deployers, management and boards, and give affected people a clear route to challenge consequential decisions.

On July 10, Malaysia's Ministry of Digital opened public consultation on a proposed Artificial Intelligence Governance Bill through the National AI Office, which has since been institutionalized as AI Malaysia. The consultation closed in early August.

The process has already moved on. On September 10, Digital Minister Gobind Singh Deo said the draft Bill had been completed, was being submitted to Cabinet, and was expected to reach Parliament this quarter or by the first quarter of 2027 at the latest.

One design choice from the consultation deserves to survive the drafting process: responsibilities should be distributed according to the role and degree of control of the parties involved. Earlier this year, Deputy Digital Minister Wilson Ugak Kumbong similarly said accountability should be traceable through a system's chain of ownership, operation and control.

That is the right starting point because consequential AI rarely has one author.

Vietnam's AI Law has been in force since March 2026, while South Korea's AI Basic Act took effect in January. Malaysia can learn from countries that moved earlier. The harder task is turning a sound principle into something that can be checked.

The chain is already fragmented

Consider a Malaysian bank using a vendor's model to flag loan applications for fraud, or a ministry using an AI system to help score applications for a subsidy.

The model may come from a global developer and run on a cloud provider's infrastructure. A systems integrator may configure it. The organization chooses the use case, adds its data, sets thresholds and decides how much weight the output carries. Employees review the result. Management sets their targets and escalation rules. The board oversees the broader risk appetite and governance structure.

When it wrongly flags or rejects someone, each participant can point elsewhere. The developer did not choose the use case. The deployer did not build the model. The reviewer followed the workflow. Management can say a human remained in the loop.

The person affected is left with a simpler question:

Who was responsible for making sure this should not happen?

The consultation framework recognizes the problem. Its proposed principle of clear accountability says responsibility should remain attributable to identifiable people and should not disappear into the AI system itself.

The gap is not the principle. It is what organizations will actually have to document and prove.

Roles should follow control, not the contract

The proposed developer-versus-deployer distinction is useful because it follows function. A developer materially shapes what a system can do; a deployer determines whether, where and how it operates, including its monitoring, suspension or withdrawal. The same organization can occupy more than one role.

That approach resembles the logic of the European Union's AI Act. Under Article 25, a deployer or other party can become the provider of a high-risk system if it rebrands it, substantially modifies it or changes its intended purpose in a way that triggers the relevant conditions.

This is where a control-based framework earns its value, and where it can fail.

Contracts can distribute cost and indemnities. They should not let a party contract away duties attached to control it actually exercises. A company that can monitor, suspend or withdraw a system should not be able to say those responsibilities disappeared because the contract used different labels.

Above that operating chain sit management and the board.

The board does not run the model. But the Securities Commission Malaysia has already warned that without sufficient capability, board oversight of AI can become superficial and overly dependent on management or external providers. Its December 2025 corporate-governance discussion paper called for clear, timely and decision-useful information, while the SC's July 2026 consultation proposed clearer expectations for PLC board oversight of technology, including AI. AI Malaysia's Boardroom Primer addresses the same governance gap from another direction.

The AI law and corporate-governance framework are converging on the same question: who has the power to act, and can that responsibility be traced?

Write the chain down before the system goes live

For a material AI system, an organization should be able to answer six questions before deployment and produce the answers afterward:

  1. Who approved this use case?
  2. Who can override or stop the system, and on what authority?
  3. Who monitors performance, bias, safety and unexpected behavior?
  4. Who reports an incident, and to whom?
  5. Who is accountable for the human review around the system?
  6. Where does an affected person go to question or challenge the outcome?

None of this requires directors to become engineers. It requires authority to be traceable. A reviewer should count as a control only if that person has the information, time and authority to disagree with the system; otherwise, the human can become a procedural signature attached to an automated decision.

Incident reporting also needs clarity. The Ministry's July press release described incident-reporting systems and safeguards as strongly encouraged, while summaries of the consultation framework describe a reporting requirement that extends to near misses. The statute should resolve that difference clearly for high-risk systems.

Malaysia's capital market already offers a useful precedent. The Securities Commission's technology-risk regime requires capital market entities to notify it of technology and cyber incidents, and its guidance expressly recognizes near-miss events.

A regulator that learns only from completed harms is learning too late.

The public-interest test is whether someone can still be held responsible

Malaysia now has an official Automated Decision-Making and Profiling Guideline within its personal-data protection framework.

But the Personal Data Protection Act itself does not apply to the Federal Government or State Governments. That creates a distinct public-sector accountability question when AI materially shapes a government decision about a person.

A horizontal AI law should not reproduce that blind spot.

For consequential systems, an affected person should be able to know that AI materially influenced a decision, obtain meaningful human reconsideration, and identify the institution accountable for the system. Public bodies should be subject to a traceable accountability framework when they deploy AI, not treated as an afterthought because a separate data-protection statute does not cover them.

What should survive into the Bill

Malaysia has chosen a strong organizing principle. The drafting now has to make it operational.

Four provisions would do much of the work:

None of this requires a longer list of principles.

It requires a shorter list of excuses.

The strongest version of Malaysia's AI framework will not be the one with the most elegant taxonomy of risk. It will be the one that makes it difficult for any organization to say, after something has gone wrong:

"That part belonged to someone else."

Source notes

Related analysis

The Model Came From a Vendor. The Accountability Did Not.

Read the companion insight →Start a conversation →